基于CT影像组学的结直肠癌肝转移与原发性肝癌病灶分类研究

Classification of colorectal liver metastases and hepatocellular carcinoma lesions based on CT-radiomics

目的探索结直肠癌肝转移(colorectal liver metastases,CRLM)与原发性肝癌(hepatocellular carcinoma,HCC)影像组学特征的差异,以实现对CRLM的精准识别。方法纳入河北大学附属医院102例经病例证实的CRLM和HCC患者术前CT增强影像,将其以7∶3的比例随机分配到训练集和测试集。首先,采用基于Python的Pyradiomics包从肝脏病灶中提取影像组学特征;然后,利用最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)和递归消除(recursive feature elimination,RFE)方法选择出最优特征集合;再应用支持向量机(support vector machine,SVM)、K-近邻(k-nearest neighbor,KNN)和随机森林(random forest,RF)、逻辑回归(logistic regression,LR)4种分类器算法训练模型,以受试者工作特征曲线下面积(the area under the receiver operating characteristic curve,AUC)、准确率、敏感度和特异度来评估4种分类器的性能。结果应用SVM分类器算法训练的模型对CRLM识别效能较高(准确率为93%,特异度为88%,灵敏度为100%,AUC值为0.94)。结论本文应用CT影像组学方法提取病灶异质性特征,并通过特征选择找到训练模型效果最佳的特征集合,应用SVM分类器算法训练的模型能够比较准确地识别出CRLM病灶,对医学诊断具有良好的应用价值。

Objective To explore the difference in radiomics characteristics between colorectal liver metastasis(CRLM)and hepatocellular carcinoma(HCC),so as to realize accurate recognition of CRLM.Methods One hundred and two preoperative CT-enhanced images of CRLM and HCC patients from the Affiliated Hospital of Hebei University are included,and they are randomly assigned to the training set and the test set at a ratio of 7∶3.First,we use the Python-based Pyradiomics package to extract radiomics features from liver lesions.Second,we use the least absolute contraction and selection operator(LASSO)and recursive elimination(RFE)methods to select the optimal feature set.Then we apply the support vector machine(SVM),K-nearest neighbor(KNN),random forest(RF),logistic regression(LR)four classifier algorithms to train models.This article uses the area under the receiver operating characteristic curve(AUC),accuracy,sensitivity and specificity to evaluate the performance of the four classifiers.Results The model trained with the SVM classifier algorithm has a high recognition performance for CRLM(accuracy rate of 93%,specificity of 88%,sensitivity of 100%,and AUC value of 0.94).Conclusions In this paper,CT-radiomics method is used to extract the heterogeneous features of the lesion,and the feature set with the best training model effect is found through feature selection.The model trained by the SVM classifier algorithm can identify CRLM lesions more accurately,which is useful for medical diagnosis.